Features of the development of GPT chats by Russian developers

In anticipation of large-scale import substitution in software, the largest Russian IT holdings are launching their own generative neural networks. They are already used by small and medium-sized businesses to optimize processes and prepare analytics. Experts note that in the future, generative AI will become even more independent, it will be able to become an operator of complex systems and will design solutions for business. Kommersant understood the specifics of developing their own GPT chats of domestic companies.

Russian matrix

Russian developers, wanting to repeat the success of ChatGPT from OpenAI, began actively launching their own analogues of the technology last year. Their networks, aimed at local audiences, respond more accurately to requests in Russian. Therefore, they are easier to adapt to the needs of local businesses.

Yandex was one of the first to create a solution similar to ChatGPT. Its “YandexGPT,” introduced in May 2023, can do what ChatGPT does. Compose text, answer questions, translate, extract videos and articles up to 30 thousand characters, write code. The neural network from Yandex was trained, like other similar neural networks, on terabytes of texts from the Internet, books and other publicly available sources. They were broken into small fragments and loaded into the model.

Yandex itself was the first to integrate the neural network into business processes. He implemented it into his services – connected it to the Alice voice assistant and Yandex.Browser. The Yandex press service explained to Kommersant that YandexGPT currently works in more than ten company services.

Yandex has another popular neural network – Masterpiece, a copy of the American Midjourney. It is capable of generating images based on text queries. The neural network was launched last spring, then it literally blew up the RuNet: hundreds of people rushed to create images in it and post them online.

The main competitor of YandexGPT is the neural network from Sberbank GigaChat. It was presented in April 2023, and already in September it opened access to everyone. The neural network can do everything that its counterpart from Yandex can do. GigaChat integrates another neural network from Sber, Kandinsky, which, like Shedevroom, specializes in creating images based on a text request. This model can draw in 20 styles and complete parts of the picture. Denis Dimitrov, scientific and technical director of Kandinsky development, told Kommersant that training of the model continues. It is planned to introduce its new versions in 2024. The main emphasis will be on the ability to generate video content.

In February of this year, two more neural networks were presented. One of them is MTS AI Chat from the MTS structure that develops AI, MTS AI. This neural network can work with text, just like its competitors from Yandex and Sber. In the near future, it is planned to teach her how to draw, make videos and write code.

MTS AI Chat is not yet publicly available: the company positions it as a development for the corporate segment. According to MTS AI senior product manager Sergei Ponomarenko, the main efforts are now aimed at improving the model’s ability to solve business problems. For example, searching for answers in corporate databases. Today, the solution has become a leader among competitors in extracting information from text, notes Sergey Ponomarenko.

Another neural network presented in February was developed by VK. It still doesn’t have a name. However, it has already been implemented in some services of the corporation, for example, in Mail.ru mail, where it can retell an incoming letter, as well as in the calendar, notes, and cloud storage. VK Vice President for AI, content and recommendation services Anton Frolov said that today the neural network can create texts on given topics (as, indeed, all its competitors), for example, write congratulations or generate ideas for posts.

Didn’t make it to the West

The power of neural networks is determined by the number of parameters they were trained on. From a mathematical point of view, a neural network is a polynomial with a large number of variables, or parameters. Her teaching is precisely about finding them. The more parameters a neural network uses, the more accurate it is.

For example, Open AI GPT-3 uses 175 billion parameters, and GigaChat from Sber – only 18 billion. Therefore, the quality of answers from the same Open AI GPT-3 will be noticeably higher.

As Alexey Zamesov, head of special projects at RU-CENTER, explains, there are giants abroad that have access to virtually unlimited computing resources. This allows them to create networks with a large number of parameters – from 70 billion or more. Russian neural networks do not yet have such capabilities.

But Russian generative AI has its strengths. Russian developments are local projects; they are taught using Russian-language texts. As Alexey Zamesov notes, Western networks learn from texts in all languages: “There are approximately 5–6% of texts in Russian on the Internet. Therefore, Western developers pay little attention to Russian-language content when training their neural networks. Accordingly, domestic solutions for Russian-language queries will be more accurate.”

This is confirmed by Vasily Krikunov, an expert in the field of AI and advanced analytics at the IT company Axenix. According to him, Russian neural networks are able to understand slang and specific terminology, which Western developments cannot always perceive. Another advantage of Russian neural networks, according to Grigory Shershukov, product director at Nanosemantics, is that they are created taking into account local legislation. This makes it easier to integrate them into business processes.

Monetization of intelligence

Russian business has already begun to actively use domestically developed generative neural networks. In particular, they are loved by SMM agencies, web studios and design studios. “If we talk about commercial use, then most often neurons today work in the design, marketing and sales industries, in call centers, and in software development. But we are only at the beginning of the journey. Products are appearing for a variety of industries, for example, for recruiting personnel,” says Alexey Zamesov. Neural networks are already helping to summarize reviews into a short summary of products presented on Yandex Market, fill out applicants’ questionnaires and evaluate their expert skills by listening to interviews with them.

Russian businesses also use neural networks to communicate with clients, analyze and prepare documents and presentations, continues Alexey Shpilman, head of the “Practice and Data” working group of the AI ​​Alliance.

Vasily Krikunov said that one of the top five Russian banks connected GigaChat as an interface to the corporate database. GigaChat is also used there in support of corporate services.

The Kandinsky neural network is used by Sberbank itself to create individual design elements in its own products, at various events and exhibitions, Denis Dimitrov said. Kandinsky also recently created designs for a series of porcelain mugs from the Imperial Porcelain Factory.

The online business service Bitrix24 has integrated YandexGPT into its AI assistant. And the robot caller Dasha.AI, created in Russia, is testing the YandexGPT API (a service for testing YandexGPT in business products) to automate voice scripts in call centers. “Recently, we have seen particular demand for automation of sales using neural networks, their compilation of answers and reviews on marketplaces. We also see a tendency for companies to integrate YandexGPT into ready-made IT products,” noted the Yandex press service.

The MTS AI Chat neural network is already used by the company’s clients to automatically compile reports on the results of meetings.

The neural network can “listen” to calls and, after the meeting, fill out a reporting form in the CRM, says Sergey Ponomarenko. This form reflects the tasks that were discussed during the video call.

VK said that for now, unlike its competitors, it does not want to promote its neural network for business. According to Anton Frolov, generative AI in the corporation was created primarily as an assistant for users and authors when creating content.

Experts note that businesses have just begun to implement neural networks, and they will be used even more widely in the future. “These developments are capable of creating vector images and 3D objects. This further expands their capabilities both for the design of new devices and assemblies, and for the creation of new drugs,” notes Alexey Shpilman. “In the future, neural networks will be able to independently launch the necessary modules for calculations and even create other AI systems.”

Similar technologies will help in production. For example, they will be able to write instructions for performing repairs, technical maps for equipment maintenance, says Vasily Krikunov. “In the field of finance, generative AI will be able to automate routine – it can process primary documentation and analyze financial reports. It will also help analysts better understand the market – quickly monitor inflation and exchange rate expectations, industry trends,” the expert concludes.

Alexander Stolyarov